Deep Learning: Automated Surface Characterization of Porous Media to Understand Geological Fluid Flow Wonjin Yun Stanford University wyun@stanford.edu Abstract In this paper, FCN and CNN were trained on the image data set that were prepared from the in-house fabricated micro-fluidic device and fluid injection experiment. Trained FCN and CNN model provide the best validation accuracy of 0.81 and 0.83, respectively, for the surface wettability classification of sandstone and carbonate pore structure. 1. Introduction Energy underpins all aspects of modern life. Energy use is directly correlated with broad measures of societal health including the human development index, and female life expectance at birth more is more. Human ingenuity, in- novation and technology has been focusing on relieved the greater pressure on worlds energy demand allowing peo- ple to gain more access to energy and to power their higher standards of living. In many of the challenges we face today as geoscientists, in particular in the context of water and energy resources, fluid invasion into a porous soil or sediment is a key process in different scale ranged from pore to reservoir scale shown in Fig.1. Examples include hydrocarbon migration and re- covery, methane venting from hydrate-bearing sediments, drying and wetting of soils, and carbon geosequestration. Complex interplay between capillary, viscous, and grav- itational forces, wettability effects, and the underlying het- erogenous pore geometry, leads to ramified, preferential flow paths.Particularly, Fig.1 demonstrate the key mech- anism at pore-scale where oil invade into narrower pore throat (R th ) when pressure difference (P 1 - P 2 ) is higher than the capillary pressure (2 * σ * ( 1 R th - 1 R b )). Here, σ interfacial tension that is strongly related with the surface wettability. P 1 - P 2 > 2 * σ * ( 1 R th - 1 R b ) Figure 1. Schematic diagram of fluid invasion into a porous soil or sediment at different scale from pore to reservoir scale. Hence, the microscale visualization of fluids in complex geometry brings better understanding of fluid movement, droplet generation and other effects, especially for the effect of wettability of the surface geometry on fluid flow.[8] The effect of wettability, simply oil-wet and water-wet shown in Fig.2 , on the fluid movement can be studied, visualized and modeled on a micro-level representative of the actual rock structure, so called micromodel. Micromodel is a silicon- based microfluidic device that are particularly useful labo- ratory tools for the direct visualization and image acquisi- tion of fluid flow revealing mechanisms controlling flow and transport phenomena in natural porous media [7][6]. Micro- scopic image in Fig.2 demonstrate fluid saturation patterns for oil or water-wet condition of sandstone-like pore geom- etry. 1.1. Challenges Predicting the emergent patterns is challenging, because of the sensitivity to pore-scale details and the large number of coupled mechanisms and governing parameters which vary over a wide range of values and scales. To evaluate the variability of multi-phase flow properties of porous media at the pore scale, it is necessary to acquire a large number of representative samples of the void-solid structure. Indeed, image analysis on microscopic images requires tremendous 1